Everyone Is Calling an AI Capex Bubble. Almost No One Agrees on How to Measure One.
The word "bubble" is now attached to AI capex on every earnings call, every cable panel, and every fund letter I read. What strikes me is that the people shouting it and the people waving it off are not actually arguing about the same number. One camp is measuring spend against profit. The other is measuring spend against demand. They both think they are winning, because they are not playing on the same board.
I cover capital for TensorFeed, and I have watched this debate harden into two scripts that rarely touch. So let me do the boring thing first: lay out the bear case and the bull case on their own terms, with real figures attributed, then show the one lens that actually travels across a century of buildouts. That last lens is the only honest way I have found to ask whether this is a mania or a buildout, and the answer is less obvious than either camp wants.
The bear case: the spend is outrunning the return
Start with the raw outlay, because it is genuinely enormous. The four largest US hyperscalers (Microsoft, Alphabet, Amazon, and Meta) spent roughly $448 billion on capex in 2025, up from about $162 billion in 2022, and have guided 2026 capex to roughly $600 to $725 billion, the bulk of it AI and datacenter, per their own guidance and a tally by Tom's Hardware and IEEE ComSoc. TrendForce puts the top nine cloud and AI infrastructure providers near $830 billion for 2026. Goldman Sachs models roughly $765 billion of AI capex in 2026 and about $7.6 trillion cumulatively from 2026 through 2031.
Now the part that worries the bears. A widely cited 2025 MIT study (Project NANDA, "The GenAI Divide") found that roughly 95 percent of enterprise generative-AI pilots showed no measurable P&L return. Set that against capex that is compressing hyperscaler free cash flow as it outruns operating cash flow, and you get the bear thesis in one sentence: the money is going out the door far faster than it is coming back.
Then there is the plumbing. Critics argue that a chunk of the demand is circular vendor financing, where the same dollars cycle between chipmaker, lab, and cloud. The examples they cite are Nvidia's reported roughly $100 billion commitment tied to OpenAI and OpenAI's reported roughly $300 billion Oracle compute deal. Investor Michael Burry has reopened bearish positions and publicly argued, as reported, that hyperscalers flatter earnings by extending the assumed useful life of AI servers, which understates depreciation. I treat that as one bear's claim rather than settled fact, but it points at a real soft spot: the accounting for how fast this hardware ages is an assumption, not a measurement.
The bull case: the demand is real and the capacity is short
The bulls do not dispute the spend. They dispute the framing of it as speculative. Their starting point is that inference demand is real and growing, not a pilot that fizzles, and that the constraint right now is supply, not interest. Operators keep reporting that GPUs and power are effectively sold out, which is not what a demand mirage looks like.
The revenue side backs that up, though I am going to keep the exact figures hedged because the private labs disclose selectively. The reported pattern is steep, recurring revenue ramps at the frontier labs and at the clouds renting them capacity. When a stranded cluster gets leased to a rival within weeks of going idle, as we documented in the SpaceX S-1 disclosure of the Anthropic-Colossus lease, that is a capacity shortage with a price, not a glut.
The bull reading of the circular-financing complaint is also worth stating fairly. A chipmaker taking equity or a supply commitment in a customer is how booms have always financed their early innings. It is risky, and it can paper over weak end demand, but vendor financing is not by itself proof of a bubble. The honest bull position is that the spend is rational if the inference economy compounds, and reckless if it does not, and that the early demand signals lean toward compounding.
Why nobody agrees: there is no shared scoreboard
Here is the core problem, and it is the reason both camps can sound right. There is no agreed scoreboard. The denominators are different, the flows are partly circular, and the depreciation is an estimate.
Take the simplest ratio, capex against return. The bears divide this year's spend by this year's realized profit and get an alarming number. The bulls divide this year's spend by a forward demand curve and get a reasonable one. Neither is lying. They have chosen different denominators, and the entire disagreement lives in that choice.
The circular flows make it worse. When Nvidia funds a lab that buys Nvidia chips through a cloud that books the revenue, the same dollar can appear as demand in three places. Counting it once or counting it three times produces wildly different pictures of how much real, independent demand exists. And the depreciation schedule, the thing Burry is poking at, is not observed at all. It is an assumption about how long a GPU stays useful, and shifting it a couple of years moves reported earnings without moving a single server. You cannot settle a bubble argument with a number that the accountants get to choose.
The one number that travels: capex as a share of GDP
When the firm-level ratios fail, I reach for the one lens that survives across very different eras: total buildout capex measured against the size of the economy. It does not care about a single company's depreciation policy or which entity books a circular dollar. It just asks how big the bet is relative to everything else being produced.
Even this number is contested, which tells you how unsettled the whole debate is. Estimates of AI capex as a share of GDP run from about 0.8 percent (Goldman, dividing total AI capex by global GDP) to roughly 2 percent (US hyperscaler capex over US GDP near $31.8 trillion). The gap is entirely the denominator: global output or US output. Goldman has noted that past technology booms peaked above 1.5 percent of GDP, so by the US-only reading the AI boom is already in historic-peak territory, and by the global reading it has room to run.
The historical analogs put both readings in perspective. The dotcom and telecom fiber buildout peaked near 1.0 to 1.2 percent of US GDP around 2000, and by 2002 only about 5 percent of the laid fiber was lit, a vivid reminder that real demand can arrive years after the capital does. Go back further and the manias get larger, not smaller: the US railroad boom ran near 4.8 percent of GNP at its 1880s peak, and the UK Railway Mania hit roughly 7 percent of GDP at its 1847 peak. So the AI buildout, even on the aggressive US-only reading near 2 percent, is bigger than dotcom but well short of the railroad manias. That single comparison reframes the question from "is this a bubble" to "which historical buildout does it rhyme with."
For readers who want to track that comparison themselves rather than take my framing on faith, TensorFeed publishes the historical capital-buildout series and the live AI capex figures at its /api/capital-cycles and /api/ai-datacenters endpoints, alongside the funding portfolio tracker. It is one of a few places to line up the railroad, fiber, and AI cycles on the same axis. The live spend, the model landscape on model wars, and the day's capital headlines on today are where I check whether the trend is bending.
Our Take
This is a buildout, not a bubble in the dotcom sense, but it is a buildout that is being financed and accounted for in ways that should make everyone nervous. Those are not contradictory statements. The demand signal is real: sold-out capacity and a rival paying nine figures a month for idle GPUs are not what fake demand looks like. At roughly 2 percent of US GDP on the aggressive reading, the bet is large but it is closer to the dotcom buildout than to the railroad manias that ran four to seven times higher. The capital is not insane relative to history.
The risk is not the size of the bet. It is the quality of the scoreboard. The 95 percent no-return pilot figure from MIT, the circular vendor financing, and the depreciation schedules that the operators get to set are three different ways of saying the same thing: the industry is marking its own homework. A buildout this large funded on assumptions this soft does not need a demand collapse to hurt people. It just needs the depreciation math to prove optimistic, or one link in the Nvidia-to-lab-to-cloud loop to stall, and the circular dollars unwind faster than they stacked up. I have watched booms that were directionally right still wipe out the investors who bought the peak, because the timing and the accounting, not the thesis, are what kill you.
So my call is specific. Stop arguing about whether it is a bubble, because the word is doing no work when the two sides cannot agree on a denominator. Watch three things instead. First, capex as a share of GDP: if the US-only reading pushes past the 1.5 percent zone where prior booms peaked and keeps climbing, the railroad comparison starts to bite. Second, the depreciation disclosures: if more operators quietly extend server useful-life assumptions the way Burry alleges, that is the tell that earnings are being managed rather than earned. Third, whether the circular financing keeps growing as a share of the demand. The thesis can be right and the trade can still be a disaster. Both of those can be true at once, and right now they are.